Image LoRA guide

How to train an image LoRA online

An image LoRA teaches a base image model a specific person, product, character, or visual style. The hard part is normally the notebook, dependencies, GPU, configuration, and artifact packaging. NiceVois keeps those steps behind one guided job.

1. Choose 8–40 useful images

Use images that show the same intended subject clearly. Variation in angle, crop, lighting, expression, and background helps the LoRA learn the subject rather than one exact picture. Avoid watermarks, corrupt files, and unrelated images.

2. Choose what the images teach

Select person, product, character, or style. This changes the caption pattern used for the dataset. Give the model a short recognizable name; NiceVois creates a unique trigger word for the package.

3. Upload once and start training

The browser uploads each image privately. NiceVois verifies the file type, dimensions, and byte count before the image queue receives the job. This queue and GPU endpoint are separate from NiceVois voice training.

4. Keep the portable output

When the run finishes, download the ZIP. It contains the LoRA .safetensors, the exact manifest, and a short guide. The package can be used outside NiceVois in tools that support the same base-model family.

Before your first run

Read how to prepare an image LoRA dataset. A smaller varied dataset is usually more useful than twenty near-identical frames.